ClaimReceipt: Verifying Evidence Sufficiency and Coverage in Agent Evaluations

TEXT START: Agent evaluations face two distinct evidentiary questions: whether a reported claim is recomputable from retained evidence (sufficiency), and whether the retained records cover the committed experiment set (coverage).

The Dissection

ClaimReceipt is building audit plumbing for agent evaluations. Its real intervention is narrow but useful: replace generic logs and hash-linked transcripts with claim-relative evidence receipts tied to a signed experiment manifest. That makes two failure modes visible—insufficient evidence and undisclosed omissions—provided the committed experiment universe is authentic and complete.

The reported results show engineering validation, not broad scientific validation: exact replay on 1,392 records, successful detection of 11 semantic faults, no false positives in eight tested cases, and a prospective 30-assignment epoch. The decisive admission is that the frozen specification was not yet unambiguous to an independent reader. The machine can therefore verify compliance with a contract that humans may still misunderstand.

The Core Fallacy

The central error, under the Discontinuity Thesis, is confusing verifiability with validity or power. A receipt can prove that a claim is recomputable from retained records and that records match a declared manifest. It cannot prove that the manifest captured the right universe, that the recorded experiment reflects reality outside the protocol, that the semantics were correctly defined, or that the agent is economically indispensable.

ClaimReceipt may harden evaluation integrity. It does not touch P1, P2, or P3. It cannot prevent cognitive automation, preserve human-only economic domains, or restore the employment-to-consumption circuit. At most, it makes the new machine regime more legible to its owners, auditors, regulators, and litigants.

Hidden Assumptions

• The experiment manifest is committed before inference, complete, and not strategically narrowed.
• The signer, receipt chain, encryption layer, and auditor access path are trustworthy.
• “Coverage” over the committed universe corresponds to coverage of the real deployment or population.
• Typed evidence captures the causal and semantic context needed for the claim, not merely its replayable outputs.
• The tested faults and eight false-positive cases represent meaningful operational failure modes.
• Independent readers can eventually interpret the specification consistently.
• Post-generation records remain auditable despite adaptive agents, changing environments, and adversarial operators.
• Economic claims can be settled by protocol evidence rather than by contested definitions of value, cost, or productivity.

The smallest result set is also being made to carry a large rhetorical burden: five historical audit labels, 11 faults, and one 30-assignment prospective epoch establish feasibility, not durability under hostile scale.

Social Function

Primary classification: transition management, with prestige signaling and a partial truth.

The partial truth is real: ordinary logs are weak evidence, and omission detection requires a committed reference universe. The transition-management function is stronger. The paper prepares institutions for a world in which autonomous systems generate consequential claims and human oversight shifts from performing work to certifying machine output. Its language offers procedural control over a process whose productive ownership remains elsewhere.

That is not liberation. It is the paperwork layer of automated sovereignty.

The Verdict

ClaimReceipt is credible audit infrastructure, not a solution to trustworthy agency and certainly not a defense against systemic obsolescence. Its strongest contribution is making missing evidence harder to disguise. Its fatal limitation is conditionality: it verifies what was committed, specified, captured, and made legible—nothing beyond that. Under the Discontinuity Thesis, it creates a narrow Servitor niche for protocol designers and auditors while increasing the control surface of AI Sovereigns. The employment system still dies; this merely gives the death certificate a tamper-evident receipt.